ARM’s Neoverse N2: Cortex A710 for Servers
chipsandcheese.com
chipsandcheese.com
back to the topic: ARM and AMD CPUs have about the same L1 instruction cache size. But don't they have a very different instruction density, giving one a noticeable advantage over the other?
Or occasionally by RISC-V "enthusiasts" who claim that none of it matters since ARM is dead and everybody is going to switch to RISC-V in the very near (but undetermined) future..
Neoverse N2 chips should have been available years ago. Once Jim Keller releases his Ascalon server RISC-V CPUs that are competitive with Zen x86 server CPUs, there is basically no room for ARM servers left since ARM couldn't manage to get their feet into the server market soon enough. So many ARM server vendors canceled their CPU projects, to the point that the safest prediction you can make about any new ARM server CPU is that it will be canceled. The thing about ARM is that it is already dead in the server market.
Not dead (yet). Just a legacy ISA.
>everybody is going to switch to RISC-V
RISC-V is inevitable.
Exactly, I find this sort of fanaticism about an instruction set (of all things) so puzzling.
Many of the comments which mention RISC-V have anything constructive to say or any meaningful arguments. All the fanboyism makes me feel a bit like I'm on reddit (or worse)...
I recognize the value of some level of coordination in marketing efforts, and as I'm no expert I just follow their recommendations, like "RISC-V is inevitable".
The A710 has a 1536-entry µop cache. That is 3/16 of the L1 cache (counting each instr as four bytes)
(Because the uop caches must fill lines aligned to real L1 cache, utilization cannot be perfect, and x86 has a somewhat harder time to achieve full utilization, while the uop cache on arm is easier to utilize because instructions are more regular, only "losing" slots when instructions are fused.)
It's not just CPUs, feels like many discussions related to personal computing in general get hijacked into Apple discussions, likely because there's many people here desperate to rationalize their purchase of mediocre hardware at a 10x markup.
The transition to ARM will happen first in the server/cloud space, but consumer will have to wait much longer for it to reach PCs, if it ever does, as X86 is also making progress.
It's not obvious to me that this is somehow inevitable. Yeah power efficiency wise Apple is ahead currently. But are there any reasons AMD/Intel why wouldn't be able to catch up and surpass Apple in a generation or two?
Eventually the gap will narrow as Apple already picked the lowest hanging fruits with the M1 design, and as we saw the jump to M2 was no longer as impressive, so we just have to wait.
It's not inconceivable that Intel might catch up and surpass TSMC. I mean they were in a somewhat similar situation they are in now back in the early to mid 2000s.
Are silly claims like this necessary? Yes, Apple massively overcharges on memory and storage upgrades because their users obviously have no other options...
But besides Macs are pretty good value if you're prioritizing battery life, slimness and fan sound.
I guess latest XPS 15/17 are fairly competitive with MBP 14/16. Especially if you don't care about battery life and heat that much, the 13700H/ 13900H seem to be on par with the M2 Pro/Max. The base price is up to $500 lower and you get replaceable RAM (also slower) and 2 m2 slots.
However I don't think there is anything that's close the the M2 Air on the market atm?
Nah, there's been a history of poor QA and just general poor build issues. Have they fixed the coil whine yet? It's still there last I checked.
I will be "that guy"; the original M1 Macbook Air struggled to compete with the Ryzen 4800u at launch. Current Ryzen mobile chips put up very well against the M2, considering the architecture gap:
https://browser.geekbench.com/macs/macbook-air-2022
https://browser.geekbench.com/v6/cpu/1604691
This is the sort of discussion the grandparent comment was lamenting though. Honest-to-god, none of this matters; dev shops will use whatever they need at the end of the day. Infecting the discussion with pedestrian "muh ARM revolution" comments always ruins it though. I like ARM for reading e-books as much as the next guy, but server use is a whole different discussion. Even mentioning Apple's consumer offerings is a punch-line like saying you want to beat OpenAI using consumer-grade GPUs. They exist in the ARM space, but they could not be any less-present in the server one. Their relationship to the topic at-hand is fraught and near non-existent if it weren't for their history of failure with Xserve.
Instruction cache sizes are not set by what you want, but by what you can get away with. The fetch path has both a lot of distance to travel and transistors to switch, and is typically a very critical path for clock speed, and while adding more pipeline stages for it helps with clocks, it also directly hurts performance near branches.
So in the end you have to limit the size depending on what clocks you target and how many stages you are willing to give the fetch path.
Windows is not really my thing for many years, so Apple is on the table in the battery life requirement, but I don’t think I’m coming back to Intel/AMD soon.
I recently tried to configure a Proliant DL385 Gen11, in order to give my sales rep a coherent list of "this is what I want" in terms of CPUs, memory and disks. And let me tell you, it's a sales experience that might just as well be a Monty Python sketch, it's so bad.
I mean literally, one of their subdomains had a certificate expired so that I had to add a security exception just to view the information. The page in question had select boxes that couldn't be scrolled, anything I did would just scroll the main page instead. Their other configurator (because ofc they need at least two different ones) was completely unusable, when you try to select a drive or something you get an explosion of recursive "to select that you also need to select this cable kit, one of those 4 cages and deselect that other thing" warnings. Like, if you know what to fix, just fix it!!
It’s clearly not the instruction set architecture. I wonder if the problem is that Arm’s business model makes it hard to justify the investment needed given the licensing fees they feel they can charge.
My reading between the lines on Nuvia was that Arm was looking to make more on Nuvia cores and providing Nuvia with substantial help. But Qualcomm broke into the party.
Happy to be told I’m wrong though.
I thought it was a good choice. They were assuming deep familiarity with modern CPU architectures through the whole text, so a one-sentence definition of all the terms would either have been too much, or not enough, depending on the reader.
I have an idea for Apple's next great product. Bring back Xserves, but with M-series processors in them this time. Then hire the Asahi Linux developers and get Linux running very solidly on them. Make it so you can buy them with either macOS or Linux pre-installed.
Sincerely,
lockhouse
Just on par with AWS graviton...
Not having to reverse engineer drivers for pretty much every single component would probably require a fraction of the effort though. And it's not like Apple would really need to actually maintain their own distro, releasing the drivers (which they are more than capabable of doing) would be perfectly sufficient.
The biggest gap right now (afaik) is Apple's virtualization limits. Regardless of the cores in your system, any Mac can host no more than 2 instances of MacOS. It's a huge kneecap to how Macs can be used as a server, and a death-blow to any serious hosting company that wants to provision them as a virtual resource.
> I would think the performance per watt would be very competitive with other server vendors on the market as well.
I can go get a 4 core Ampere ARM instance for free. The market goes so much cheaper than Apple's CPUs it's not even funny. Performance-wise, Apple kinda got their lunch ate by Nvidia's Grace server. There isn't really a demand to fill for high-performance ARM computing, and the cheap end goes so cheap that Apple can't even touch them.
I see two ways for Apple to escape this:
1. They can double-down on their ecosystem investments to create a "CUDA-killer" and accompanying hardware, spending tens-of-billions of dollars across multiple years to obsolete Nvidia's industry ties.
2. Initiate or join an Open Source pact to write a CUDA replacement or translator a-la Vulkan/DXVK that gives vendors no choice but to abandon Nvidia.
Of course, the easier play is to just pretend all of this is beneath you and continue building your current ecosystem. It's what I would do, and I don't see a rational reason why Tim Cook would read your pitch and greenlight a project like that.
Fairly good deal, the only problem being that it's Oracle.
ps. if you do decide to give them a try, be sure to upgrade to "pay-as-you-go" (and never go beyond the free offering) otherwise they will close your account without any warning (ask me how I know).
That being said, I wouldn't trust the server as far as I could throw it. It's a bad deal unless you have ephemeral compute requirements that don't matter too much.
Why not? Even just Apple's internal expense could justify it - for the same reason Google, Microsoft, Amazon etc all dog food their own cloud. Apple currently pays AWS, Akamai, Cloudflare, etc lots of $$$.
As the current focus is to expand on services and managed offerings you would focus on controlling such a cost (as is typical Apple).
> I can go get a 4 core Ampere ARM instance for free. The market goes so much cheaper than Apple's CPUs it's not even funny.
Is the market full of people that only use 4 cores? If that was really the case AWS would have closed shop already instead of being the market leader. It's not like Oracle doesn't charge you for anything else.
> Apple kinda got their lunch ate by Nvidia's Grace server
And where is Grace? Which cloud currently? It doesn't work that way. If all you think about a server and its ecosystem is hardware then lots of Android phones have better hardware than iphone e.g. in screens, battery life, etc. Was that really the deciding factor? No.
I mean by your logic, AMD's EPYC has been winning Intel Xeon's by years. Has Xeon been decimated? No. There's still room for both. AMD can't even produce enough. Considering Apple can exclusive lock TSMC down for a year on bleeding edge, who can really compete if Apple was in?
Ampere CPUs on Oracle, Hetzner and (I assume?) Azure are still remarkable cheap compared to Intel/AMD if they are sufficient for your workflows.
I'm not sure if the margins in this market are high enough for Apple to enter it and they would probably still have issues scaling production capacity if they are serious about it.
Hetzner has always been there. It didn't impact Intel/AMD use cases. There are lots of cheap providers (since the start of time) and I doubt that's the space Apple would ever get into.
So... who would they be selling to, now? I genuinely think such a "server Mac" would echo the Xserve without a fundamental shift in the way Apple treats the software community.
Why? And why is it only about hardware? So AWS, GCP and Azure all run Intel/AMD, who do they need to dethrone?
For all it matters Apple "cloud" could even use these nVidia servers themselves. Not like AWS only uses Graviton. So, so?
> So... who would they be selling to, now? I genuinely think such a "server Mac" would echo the Xserve without a fundamental shift in the way Apple treats the software community.
Right, why would they only do hardware? As is typical Apple the $ is in the ecosystem and services. The competition would be to offer cloud services that has a better experience just like they do for phones and others. The macbook pro was always more expensive and yet people still purchase them.
You are just back around to describing Xserve again. Regardless of whether it is software or hardware-based, an Apple Cloud has no upsell in today's market. That's why Xserve died, and it's why it would die again.
Which is both less relevant and quite different for enterprise services than consumer products. It's also not that obvious that Apple has a lot of expertise in this area and could provide something better than AWS/Azure/etc.. and Apple is not exactly the company which is willing to compete on lower margins.
I would actually argue that is the easy part. Apple already has access to cutting-edge nodes if they want to pay for them. The hard part is designing a server chip from Apple Silicon, a decidedly mobile SOC.
For instance; what if Apple wanted to ship a system that competes with a DGX GH200? Where would you even start with something like the M2 Ultra?
- You would either need more cores per SOC, or more SOCs per chipset to match a DGX system
- Custom memory controllers are a must (running them at regular speeds depends on how much bandwidth is left)
- PCI would need to exposed on-hardware and supported in MacOS, Linux and third-party software
- An multi-instance interconnect like Infiniband would need to be working and attractive (as well as a CUDA translator, if they want any hope)
And that's just what they would need to match the high-performance ARM industry right now. If they wanted to compete or knock Nvidia's block off, it would take a lot more investment, time and attitude-changing.
There are?
Especially after being selected for EU super computer effort.
Isn't it years behind ARM/x86 performance wise?
> nonsense to dismiss it as only being for embedded
They unlikely to be competitive anywhere else without significant investment.
> but it can also scale up to "real" compute
Did anyone doubt that you could theoretically scale up ARM to be competitive with x86 back in ~2000? It still took them several decades to achieve that..
I don't expect them to open source their processors, but they do definitely contribute back to the community. E.g. they open sourced their first risc-v vector implementation (ocelot) on top of BOOM. It's in order and they used it as a reference for their later out of order commercial designs, but they also could've keep it to them self's.
RISC-V openness has nothing to do with processor implementation, it's solely in the instruction set.
The advantage it has is that it provides compatibility across vendors and anyone can become a vendor of RISC-V IP. Comercial IP doesn't affect that, and if anything makes multiple IP vendors more likely.
The availability of server processors depends on the market rather than any technical limitation. If someone decides to take the risk of investing in the development of a strong server product then we'll get it. It depends on some company believing there is a market for it.
Just because there is a new ISA does not mean a company cannot design a modern CPU.
For example, the Vision Star Five is uses an in order CPU core that is competitive with the out of order A72 core. The performance gap on day one is much smaller. It also has UEFI support and opensource GPU drivers for Power VR are in development by imagination, so the software support will also quickly catch up or surpass ARM SBCs.
The same will happen in the server market. Nobody who works on a RISCV server chip has to wait for ARM to deliver a CPU core that doesn't suck.
Or that they show up in computer shops for regular buyers, instead of computer enthusiasts, used to install GNU/Linux into any kind of motherboard with a CPU on it.
Cause why not.. having more architectures (for no apparent reasons) sounds cool.
Likewise Apple also had their own Linux distribution that went nowhere, MkLinux.
MkLinux (MicroKernel Linux) was not a distribution but a modified Linux kernel to be loaded as a Mach server to run under CMU (later OSF) Mach 3. It abstracted idiosyncrasises of the Apple NuBus and peripherals away from the Linux kernel that did not (or could not altogether) support – the NuBus itself and the NuBus attached devices.
The distribution was a stock RedHat v4 or v5.
Amusingly, despite CMU Mach 3 being a µ-kernel, there was nothing «micro» about the size of its own binary, and – off the top of my head – the Mach binary size was larger than that of the vmlinuz at the time.
If anything, it was actually pity it didn't took off, as its design was quite interesting.
What does any of this have to do with Neoverse?